How To Give Claude Unlimited Usage

Summary

The video explains how Theo used long-running Claude agent loops to tackle unusually large projects—not by giving Claude unlimited usage in a literal sense, but by combining subscription accounts, automated success checks, and elastic compute. He emphasizes that these workflows are expensive and chaotic, and not recommended as a practical approach for most developers.

Projects and results

• TS Rust: A Rust-based replacement for parts of the TypeScript compiler, intended to speed up builds. Theo says it reached roughly 4–10× faster compilation in relevant use cases. He estimates the work represented about $38,000–$40,000 in API-equivalent model usage, though his subscription-based spend was far lower—roughly $500–$800 by his estimate.
• Lust: A custom compiler and Cargo replacement created to reduce TS Rust’s slow build cycle. Theo says one build fell from 17 minutes to 12 seconds. It is specialized for TS Rust and is not presented as a general-purpose compiler.
• Skate 3 Bench: A browser-based skateboarding game and engine made from scratch by Claude, inspired by EA’s Skate. It uses a Three.js-based engine and joystick flick controls for tricks. Theo describes the physics and gameplay as surprisingly convincing, though imperfect. The project ran for about four days, used an estimated 12 billion tokens and nearly $5,000 in inference, and required only around 25 direct prompts—many of them simply “continue.”

The core method: “hill climbing”

Theo’s central idea is to give an agent a goal it can repeatedly work toward, along with a way to judge whether each attempt is better:

1. Choose a tedious, ambitious task. Agents can persist through work that people may find repetitive or draining.
2. Define verification. The agent needs a programmatic pass/fail check, or at least another agent that can assess results. For TS Rust, this meant comparing compiler outputs and testing injected errors. For the skate game, Claude captured screenshots, compared them with reference footage, and used LLM judges to suggest improvements.
3. Provide the necessary tools and environment. Depending on the task, agents may need machines, browsers, build systems, Blender, or other software.
4. Ask what is blocking progress, then address it. Theo treats this as an ongoing loop: identify bottlenecks, add resources or tooling, and repeat.

He connects this approach to reinforcement learning with verifiable rewards (RLVR): models can improve at tasks when success and failure are checkable. His broader point is that agents perform better when the task resembles a repeated, measurable test-and-improve cycle than when they are simply asked to produce a result once.

Compute and tools

• Blacksmith: Theo used it to offload build and test jobs after local machines became a bottleneck. He reports that queue delays fell substantially, while clarifying that his intensive agent-driven use differs from typical CI workloads.
• Boat.dev: A service for quickly creating disposable virtual machines and snapshots. Theo says agents could spin up isolated environments in parallel, reducing a build-and-check cycle from as long as 40–70 minutes to about six minutes in his setup. The value, in his view, is scaling compute up and down as agent needs change.
• Browserbase: The sponsor segment presents it as managed browser infrastructure for agents, including web navigation, data retrieval, and practical browser tasks such as forms and file transfers. Theo argues that reliable access to web tools can matter as much as model capability for personal agents.

Takeaways and caveats

Theo frames agent work less as reviewing every line of code and more as managing the system around the agents: set clear objectives, make progress measurable, supply tools, and keep removing bottlenecks.

He also warns that this style can consume substantial compute and subscription capacity, create technical headaches, and encourage developers to chase ever-larger infrastructure projects. The video is an account of experimentation, not a step-by-step recommendation to replicate his spending or setup.

Main speakers and sources

• Theo, the video’s host and primary speaker, describing his projects, methods, and results.
• Claude and other LLMs, used as coding agents and, in the skate-game experiment, evaluators.
• Browserbase, featured in the sponsored segment; Blacksmith and Boat.dev are also discussed as compute tools.